1: \begin{abstract}

2: We study the problem of low-rank tensor factorization in the presence of  

3: missing data. 

4: We ask the following question: how many sampled entries do we need, to 

5: efficiently and exactly reconstruct a  tensor 

6: with a low-rank orthogonal decomposition?  

7: We propose a novel alternating minimization based method 

8: which iteratively refines estimates of the singular vectors. 

9: We show that under certain standard assumptions, 

10: our method can recover a three-mode $n\times n\times n$ dimensional rank-$r$  tensor exactly 

11: from $O(n^{3/2} r^5 \log^4 n)$ randomly sampled entries. 

12: In the process of proving this result, we 

13: solve two challenging sub-problems for tensors with missing data. 

14: First, in the process of analyzing the initialization step, 

15: we  prove a generalization of a celebrated result by Szemer\'edie et al. on the spectrum of random graphs. 

16: Next,  we prove global convergence of alternating minimization with a good initialization. 

17: Simulations suggest that the dependence of the sample size on dimensionality $n$ is indeed tight. 

18: \end{abstract}

19: